کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Edge Deployment of Quantized Transformer Models for Remaining Useful Life Prediction

نویسندگان :

Reza Adinepour¹ Shayan Naghizadeh² Morteza Saheb Zamani³

1- دانشگاه صنعتی امیرکبیر(پلی تکنیک تهران) 2- دانشگاه صنعتی امیرکبیر(پلی تکنیک تهران) 3- دانشگاه صنعتی امیرکبیر(پلی تکنیک تهران)

کلمات کلیدی :

Remaining Useful Life،Predictive Maintenance،Transformer Network،Edge Computing،Quantization

چکیده :

Predictive maintenance increasingly relies on the estimation of remaining useful life (RUL) at the edge, where computation and power budgets are severely constrained. Transformer models provide state-of-the-art accuracy for RUL prediction, but their deployment on embedded platforms remains challenging due to high memory consumption and latency. This paper presents an integrated optimization pipeline that enables efficient deployment of a Transformer-based RUL model on a Raspberry Pi device. First, a hybrid 8-bit quantization scheme reduces the model weight memory footprint by a factor of four while maintaining predictive accuracy. Then, a hardware-aware C++ inference engine is developed using ARM NEON SIMD intrinsics to compensate for the performance loss introduced by quantization. The final system achieves a latency of approximately 25.2 ms per sequence, which is close to the floating-point baseline, while the coefficient of determination remains around 0.991 on the XJTU-SY bearing dataset. These results demonstrate that high-capacity Transformer models can be made practical for real-time industrial edge applications.